Performance Comparison of EfficientNet-B5, B6, and B7 for Rice Plant Disease Classification

Rapid and accurate paddy disease identification is essential for food security. This study compares EfficientNetB5, B6, and B7 for classifying paddy leaf diseases using transfer learning with the native squeeze-and-excitation blocks inherent to the EfficientNet architecture. The dataset comprised ten disease categories with 2,077 test samples, partitioned into 70% training, 15% validation, and 15% testing using stratified random sampling. EfficientNetB5 achieved the highest accuracy of 94.46%, outperforming B6 (93.89%) and B7 (91.09%). The macro average F1-score for B5 was 93.65%, followed by B6 (92.89%) and B7 (88.94%). Confusion matrices revealed B5 excelled at distinguishing visually similar symptoms, especially for bacterial leaf streak and dead heart. B7 showed the lowest performance due to overfitting from increased model complexity. Training and loss curves indicated B5 achieved the most stable convergence with minimal gap between training and validation metrics. Statistical significance testing using McNemar's test confirmed that B5's performance improvement over B7 was significant (p < 0.05), while the difference between B5 and B6 was not statistically significant. These findings prove that larger model size does not guarantee better classification performance for paddy disease recognition. EfficientNetB5 provides an optimal balance between computational efficiency and predictive accuracy